Evolution of Molecular Representations

Recording

Learning Objectives

• Read and write SMILES, InChI, and SELFIES strings for simple organic molecules
• Compute Morgan (ECFP) fingerprints and understand the radius and bit-length hyperparameters
• Compare expert-crafted descriptors, fingerprints, and learned graph-based representations
• Choose an appropriate representation for a given property-prediction task

Key Takeaways

• Takeaway 1. SMILES are compact and human-readable but encode the same molecule in multiple ways — always canonicalise before featurising to avoid data leakage.

• Takeaway 2. Morgan fingerprints (ECFP) remain a strong baseline: they are interpretable, fast to compute, and perform well on small datasets where neural representations overfit.

• Takeaway 3. Representations are not neutral — they embed assumptions about which structural features matter. A fingerprint that ignores 3-D geometry will fail on conformer-sensitive properties.

• Takeaway 4. SELFIES guarantee syntactic validity by construction, making them the preferred string encoding for generative models where invalid SMILES outputs would break the pipeline.

Further Reading & Resources